ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/libecb/ecb.h
(Generate patch)

Comparing libecb/ecb.h (file contents):
Revision 1.159 by sf-exg, Sun Feb 22 16:35:20 2015 UTC vs.
Revision 1.198 by root, Sat Jul 31 16:13:30 2021 UTC

1/* 1/*
2 * libecb - http://software.schmorp.de/pkg/libecb 2 * libecb - http://software.schmorp.de/pkg/libecb
3 * 3 *
4 * Copyright (©) 2009-2015 Marc Alexander Lehmann <libecb@schmorp.de> 4 * Copyright (©) 2009-2015,2018-2021 Marc Alexander Lehmann <libecb@schmorp.de>
5 * Copyright (©) 2011 Emanuele Giaquinta 5 * Copyright (©) 2011 Emanuele Giaquinta
6 * All rights reserved. 6 * All rights reserved.
7 * 7 *
8 * Redistribution and use in source and binary forms, with or without modifica- 8 * Redistribution and use in source and binary forms, with or without modifica-
9 * tion, are permitted provided that the following conditions are met: 9 * tion, are permitted provided that the following conditions are met:
… …
40 40
41#ifndef ECB_H 41#ifndef ECB_H
42#define ECB_H 42#define ECB_H
43 43
44/* 16 bits major, 16 bits minor */ 44/* 16 bits major, 16 bits minor */
45#define ECB_VERSION 0x00010004 45#define ECB_VERSION 0x00010009
46 46
47#ifdef _WIN32 47#include <string.h> /* for memcpy */
48
49#if defined (_WIN32) && !defined (__MINGW32__)
48 typedef signed char int8_t; 50 typedef signed char int8_t;
49 typedef unsigned char uint8_t; 51 typedef unsigned char uint8_t;
52 typedef signed char int_fast8_t;
53 typedef unsigned char uint_fast8_t;
50 typedef signed short int16_t; 54 typedef signed short int16_t;
51 typedef unsigned short uint16_t; 55 typedef unsigned short uint16_t;
56 typedef signed int int_fast16_t;
57 typedef unsigned int uint_fast16_t;
52 typedef signed int int32_t; 58 typedef signed int int32_t;
53 typedef unsigned int uint32_t; 59 typedef unsigned int uint32_t;
60 typedef signed int int_fast32_t;
61 typedef unsigned int uint_fast32_t;
54 #if __GNUC__ 62 #if __GNUC__
55 typedef signed long long int64_t; 63 typedef signed long long int64_t;
56 typedef unsigned long long uint64_t; 64 typedef unsigned long long uint64_t;
57 #else /* _MSC_VER || __BORLANDC__ */ 65 #else /* _MSC_VER || __BORLANDC__ */
58 typedef signed __int64 int64_t; 66 typedef signed __int64 int64_t;
59 typedef unsigned __int64 uint64_t; 67 typedef unsigned __int64 uint64_t;
60 #endif 68 #endif
69 typedef int64_t int_fast64_t;
70 typedef uint64_t uint_fast64_t;
61 #ifdef _WIN64 71 #ifdef _WIN64
62 #define ECB_PTRSIZE 8 72 #define ECB_PTRSIZE 8
63 typedef uint64_t uintptr_t; 73 typedef uint64_t uintptr_t;
64 typedef int64_t intptr_t; 74 typedef int64_t intptr_t;
65 #else 75 #else
… …
67 typedef uint32_t uintptr_t; 77 typedef uint32_t uintptr_t;
68 typedef int32_t intptr_t; 78 typedef int32_t intptr_t;
69 #endif 79 #endif
70#else 80#else
71 #include <inttypes.h> 81 #include <inttypes.h>
72 #if UINTMAX_MAX > 0xffffffffU 82 #if (defined INTPTR_MAX ? INTPTR_MAX : ULONG_MAX) > 0xffffffffU
73 #define ECB_PTRSIZE 8 83 #define ECB_PTRSIZE 8
74 #else 84 #else
75 #define ECB_PTRSIZE 4 85 #define ECB_PTRSIZE 4
76 #endif 86 #endif
77#endif 87#endif
78 88
79#define ECB_GCC_AMD64 (__amd64 || __amd64__ || __x86_64 || __x86_64__) 89#define ECB_GCC_AMD64 (__amd64 || __amd64__ || __x86_64 || __x86_64__)
80#define ECB_MSVC_AMD64 (_M_AMD64 || _M_X64) 90#define ECB_MSVC_AMD64 (_M_AMD64 || _M_X64)
91
92#ifndef ECB_OPTIMIZE_SIZE
93 #if __OPTIMIZE_SIZE__
94 #define ECB_OPTIMIZE_SIZE 1
95 #else
96 #define ECB_OPTIMIZE_SIZE 0
97 #endif
98#endif
81 99
82/* work around x32 idiocy by defining proper macros */ 100/* work around x32 idiocy by defining proper macros */
83#if ECB_GCC_AMD64 || ECB_MSVC_AMD64 101#if ECB_GCC_AMD64 || ECB_MSVC_AMD64
84 #if _ILP32 102 #if _ILP32
85 #define ECB_AMD64_X32 1 103 #define ECB_AMD64_X32 1
86 #else 104 #else
87 #define ECB_AMD64 1 105 #define ECB_AMD64 1
88 #endif 106 #endif
107#endif
108
109#if ECB_PTRSIZE >= 8 || ECB_AMD64_X32
110 #define ECB_64BIT_NATIVE 1
111#else
112 #define ECB_64BIT_NATIVE 0
89#endif 113#endif
90 114
91/* many compilers define _GNUC_ to some versions but then only implement 115/* many compilers define _GNUC_ to some versions but then only implement
92 * what their idiot authors think are the "more important" extensions, 116 * what their idiot authors think are the "more important" extensions,
93 * causing enormous grief in return for some better fake benchmark numbers. 117 * causing enormous grief in return for some better fake benchmark numbers.
… …
115 #define ECB_CLANG_EXTENSION(x) 0 139 #define ECB_CLANG_EXTENSION(x) 0
116#endif 140#endif
117 141
118#define ECB_CPP (__cplusplus+0) 142#define ECB_CPP (__cplusplus+0)
119#define ECB_CPP11 (__cplusplus >= 201103L) 143#define ECB_CPP11 (__cplusplus >= 201103L)
144#define ECB_CPP14 (__cplusplus >= 201402L)
145#define ECB_CPP17 (__cplusplus >= 201703L)
120 146
121#if ECB_CPP 147#if ECB_CPP
122 #define ECB_C 0 148 #define ECB_C 0
123 #define ECB_STDC_VERSION 0 149 #define ECB_STDC_VERSION 0
124#else 150#else
… …
126 #define ECB_STDC_VERSION __STDC_VERSION__ 152 #define ECB_STDC_VERSION __STDC_VERSION__
127#endif 153#endif
128 154
129#define ECB_C99 (ECB_STDC_VERSION >= 199901L) 155#define ECB_C99 (ECB_STDC_VERSION >= 199901L)
130#define ECB_C11 (ECB_STDC_VERSION >= 201112L) 156#define ECB_C11 (ECB_STDC_VERSION >= 201112L)
157#define ECB_C17 (ECB_STDC_VERSION >= 201710L)
131 158
132#if ECB_CPP 159#if ECB_CPP
133 #define ECB_EXTERN_C extern "C" 160 #define ECB_EXTERN_C extern "C"
134 #define ECB_EXTERN_C_BEG ECB_EXTERN_C { 161 #define ECB_EXTERN_C_BEG ECB_EXTERN_C {
135 #define ECB_EXTERN_C_END } 162 #define ECB_EXTERN_C_END }
… …
150 177
151#if ECB_NO_SMP 178#if ECB_NO_SMP
152 #define ECB_MEMORY_FENCE do { } while (0) 179 #define ECB_MEMORY_FENCE do { } while (0)
153#endif 180#endif
154 181
182/* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/compiler_ref/compiler_builtins.html */
183#if __xlC__ && ECB_CPP
184 #include <builtins.h>
185#endif
186
187#if 1400 <= _MSC_VER
188 #include <intrin.h> /* fence functions _ReadBarrier, also bit search functions _BitScanReverse */
189#endif
190
155#ifndef ECB_MEMORY_FENCE 191#ifndef ECB_MEMORY_FENCE
156 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 192 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110
193 #define ECB_MEMORY_FENCE_RELAXED __asm__ __volatile__ ("" : : : "memory")
157 #if __i386 || __i386__ 194 #if __i386 || __i386__
158 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory") 195 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory")
159 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 196 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
160 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 197 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("" : : : "memory")
161 #elif ECB_GCC_AMD64 198 #elif ECB_GCC_AMD64
162 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") 199 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory")
163 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 200 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
164 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 201 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("" : : : "memory")
165 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ 202 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__
166 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") 203 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory")
204 #elif defined __ARM_ARCH_2__ \
205 || defined __ARM_ARCH_3__ || defined __ARM_ARCH_3M__ \
206 || defined __ARM_ARCH_4__ || defined __ARM_ARCH_4T__ \
207 || defined __ARM_ARCH_5__ || defined __ARM_ARCH_5E__ \
208 || defined __ARM_ARCH_5T__ || defined __ARM_ARCH_5TE__ \
209 || defined __ARM_ARCH_5TEJ__
210 /* should not need any, unless running old code on newer cpu - arm doesn't support that */
167 #elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \ 211 #elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \
168 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ 212 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ \
213 || defined __ARM_ARCH_6T2__
169 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory") 214 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory")
170 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ 215 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \
171 || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ 216 || defined __ARM_ARCH_7R__ || defined __ARM_ARCH_7M__
172 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") 217 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory")
173 #elif __aarch64__ 218 #elif __aarch64__
174 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory") 219 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory")
175 #elif (__sparc || __sparc__) && !__sparcv8 220 #elif (__sparc || __sparc__) && !(__sparc_v8__ || defined __sparcv8)
176 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") 221 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory")
177 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") 222 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory")
178 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") 223 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore")
179 #elif defined __s390__ || defined __s390x__ 224 #elif defined __s390__ || defined __s390x__
180 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") 225 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory")
… …
203 #if ECB_GCC_VERSION(4,7) 248 #if ECB_GCC_VERSION(4,7)
204 /* see comment below (stdatomic.h) about the C11 memory model. */ 249 /* see comment below (stdatomic.h) about the C11 memory model. */
205 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) 250 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST)
206 #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE) 251 #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE)
207 #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE) 252 #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE)
253 #undef ECB_MEMORY_FENCE_RELAXED
254 #define ECB_MEMORY_FENCE_RELAXED __atomic_thread_fence (__ATOMIC_RELAXED)
208 255
209 #elif ECB_CLANG_EXTENSION(c_atomic) 256 #elif ECB_CLANG_EXTENSION(c_atomic)
210 /* see comment below (stdatomic.h) about the C11 memory model. */ 257 /* see comment below (stdatomic.h) about the C11 memory model. */
211 #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) 258 #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST)
212 #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE) 259 #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE)
213 #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE) 260 #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE)
261 #undef ECB_MEMORY_FENCE_RELAXED
262 #define ECB_MEMORY_FENCE_RELAXED __c11_atomic_thread_fence (__ATOMIC_RELAXED)
214 263
215 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ 264 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__
216 #define ECB_MEMORY_FENCE __sync_synchronize () 265 #define ECB_MEMORY_FENCE __sync_synchronize ()
217 #elif _MSC_VER >= 1500 /* VC++ 2008 */ 266 #elif _MSC_VER >= 1500 /* VC++ 2008 */
218 /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */ 267 /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */
… …
228 #elif defined _WIN32 277 #elif defined _WIN32
229 #include <WinNT.h> 278 #include <WinNT.h>
230 #define ECB_MEMORY_FENCE MemoryBarrier () /* actually just xchg on x86... scary */ 279 #define ECB_MEMORY_FENCE MemoryBarrier () /* actually just xchg on x86... scary */
231 #elif __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 280 #elif __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110
232 #include <mbarrier.h> 281 #include <mbarrier.h>
233 #define ECB_MEMORY_FENCE __machine_rw_barrier () 282 #define ECB_MEMORY_FENCE __machine_rw_barrier ()
234 #define ECB_MEMORY_FENCE_ACQUIRE __machine_r_barrier () 283 #define ECB_MEMORY_FENCE_ACQUIRE __machine_acq_barrier ()
235 #define ECB_MEMORY_FENCE_RELEASE __machine_w_barrier () 284 #define ECB_MEMORY_FENCE_RELEASE __machine_rel_barrier ()
285 #define ECB_MEMORY_FENCE_RELAXED __compiler_barrier ()
236 #elif __xlC__ 286 #elif __xlC__
237 #define ECB_MEMORY_FENCE __sync () 287 #define ECB_MEMORY_FENCE __sync ()
238 #endif 288 #endif
239#endif 289#endif
240 290
241#ifndef ECB_MEMORY_FENCE 291#ifndef ECB_MEMORY_FENCE
242 #if ECB_C11 && !defined __STDC_NO_ATOMICS__ 292 #if ECB_C11 && !defined __STDC_NO_ATOMICS__
243 /* we assume that these memory fences work on all variables/all memory accesses, */ 293 /* we assume that these memory fences work on all variables/all memory accesses, */
244 /* not just C11 atomics and atomic accesses */ 294 /* not just C11 atomics and atomic accesses */
245 #include <stdatomic.h> 295 #include <stdatomic.h>
246 /* Unfortunately, neither gcc 4.7 nor clang 3.1 generate any instructions for */
247 /* any fence other than seq_cst, which isn't very efficient for us. */
248 /* Why that is, we don't know - either the C11 memory model is quite useless */
249 /* for most usages, or gcc and clang have a bug */
250 /* I *currently* lean towards the latter, and inefficiently implement */
251 /* all three of ecb's fences as a seq_cst fence */
252 /* Update, gcc-4.8 generates mfence for all c++ fences, but nothing */
253 /* for all __atomic_thread_fence's except seq_cst */
254 #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_seq_cst) 296 #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_seq_cst)
297 #define ECB_MEMORY_FENCE_ACQUIRE atomic_thread_fence (memory_order_acquire)
298 #define ECB_MEMORY_FENCE_RELEASE atomic_thread_fence (memory_order_release)
255 #endif 299 #endif
256#endif 300#endif
257 301
258#ifndef ECB_MEMORY_FENCE 302#ifndef ECB_MEMORY_FENCE
259 #if !ECB_AVOID_PTHREADS 303 #if !ECB_AVOID_PTHREADS
… …
279 323
280#if !defined ECB_MEMORY_FENCE_RELEASE && defined ECB_MEMORY_FENCE 324#if !defined ECB_MEMORY_FENCE_RELEASE && defined ECB_MEMORY_FENCE
281 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE 325 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE
282#endif 326#endif
283 327
328#if !defined ECB_MEMORY_FENCE_RELAXED && defined ECB_MEMORY_FENCE
329 #define ECB_MEMORY_FENCE_RELAXED ECB_MEMORY_FENCE /* very heavy-handed */
330#endif
331
284/*****************************************************************************/ 332/*****************************************************************************/
285 333
286#if ECB_CPP 334#if ECB_CPP
287 #define ecb_inline static inline 335 #define ecb_inline static inline
288#elif ECB_GCC_VERSION(2,5) 336#elif ECB_GCC_VERSION(2,5)
… …
352 #define ecb_deprecated __declspec (deprecated) 400 #define ecb_deprecated __declspec (deprecated)
353#else 401#else
354 #define ecb_deprecated ecb_attribute ((__deprecated__)) 402 #define ecb_deprecated ecb_attribute ((__deprecated__))
355#endif 403#endif
356 404
357#if __MSC_VER >= 1500 405#if _MSC_VER >= 1500
358 #define ecb_deprecated_message(msg) __declspec (deprecated (msg)) 406 #define ecb_deprecated_message(msg) __declspec (deprecated (msg))
359#elif ECB_GCC_VERSION(4,5) 407#elif ECB_GCC_VERSION(4,5)
360 #define ecb_deprecated_message(msg) ecb_attribute ((__deprecated__ (msg)) 408 #define ecb_deprecated_message(msg) ecb_attribute ((__deprecated__ (msg))
361#else 409#else
362 #define ecb_deprecated_message(msg) ecb_deprecated 410 #define ecb_deprecated_message(msg) ecb_deprecated
… …
371#define ecb_unused ecb_attribute ((__unused__)) 419#define ecb_unused ecb_attribute ((__unused__))
372#define ecb_const ecb_attribute ((__const__)) 420#define ecb_const ecb_attribute ((__const__))
373#define ecb_pure ecb_attribute ((__pure__)) 421#define ecb_pure ecb_attribute ((__pure__))
374 422
375#if ECB_C11 || __IBMC_NORETURN 423#if ECB_C11 || __IBMC_NORETURN
376 /* http://pic.dhe.ibm.com/infocenter/compbg/v121v141/topic/com.ibm.xlcpp121.bg.doc/language_ref/noreturn.html */ 424 /* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/language_ref/noreturn.html */
377 #define ecb_noreturn _Noreturn 425 #define ecb_noreturn _Noreturn
378#elif ECB_CPP11 426#elif ECB_CPP11
379 #define ecb_noreturn [[noreturn]] 427 #define ecb_noreturn [[noreturn]]
380#elif _MSC_VER >= 1200 428#elif _MSC_VER >= 1200
381 /* http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx */ 429 /* http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx */
… …
418#else 466#else
419 ecb_function_ ecb_const int ecb_ctz32 (uint32_t x); 467 ecb_function_ ecb_const int ecb_ctz32 (uint32_t x);
420 ecb_function_ ecb_const int 468 ecb_function_ ecb_const int
421 ecb_ctz32 (uint32_t x) 469 ecb_ctz32 (uint32_t x)
422 { 470 {
471#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
472 unsigned long r;
473 _BitScanForward (&r, x);
474 return (int)r;
475#else
423 int r = 0; 476 int r = 0;
424 477
425 x &= ~x + 1; /* this isolates the lowest bit */ 478 x &= ~x + 1; /* this isolates the lowest bit */
426 479
427#if ECB_branchless_on_i386 480#if ECB_branchless_on_i386
… …
437 if (x & 0xff00ff00) r += 8; 490 if (x & 0xff00ff00) r += 8;
438 if (x & 0xffff0000) r += 16; 491 if (x & 0xffff0000) r += 16;
439#endif 492#endif
440 493
441 return r; 494 return r;
495#endif
442 } 496 }
443 497
444 ecb_function_ ecb_const int ecb_ctz64 (uint64_t x); 498 ecb_function_ ecb_const int ecb_ctz64 (uint64_t x);
445 ecb_function_ ecb_const int 499 ecb_function_ ecb_const int
446 ecb_ctz64 (uint64_t x) 500 ecb_ctz64 (uint64_t x)
447 { 501 {
502#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
503 unsigned long r;
504 _BitScanForward64 (&r, x);
505 return (int)r;
506#else
448 int shift = x & 0xffffffffU ? 0 : 32; 507 int shift = x & 0xffffffff ? 0 : 32;
449 return ecb_ctz32 (x >> shift) + shift; 508 return ecb_ctz32 (x >> shift) + shift;
509#endif
450 } 510 }
451 511
452 ecb_function_ ecb_const int ecb_popcount32 (uint32_t x); 512 ecb_function_ ecb_const int ecb_popcount32 (uint32_t x);
453 ecb_function_ ecb_const int 513 ecb_function_ ecb_const int
454 ecb_popcount32 (uint32_t x) 514 ecb_popcount32 (uint32_t x)
… …
462 } 522 }
463 523
464 ecb_function_ ecb_const int ecb_ld32 (uint32_t x); 524 ecb_function_ ecb_const int ecb_ld32 (uint32_t x);
465 ecb_function_ ecb_const int ecb_ld32 (uint32_t x) 525 ecb_function_ ecb_const int ecb_ld32 (uint32_t x)
466 { 526 {
527#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
528 unsigned long r;
529 _BitScanReverse (&r, x);
530 return (int)r;
531#else
467 int r = 0; 532 int r = 0;
468 533
469 if (x >> 16) { x >>= 16; r += 16; } 534 if (x >> 16) { x >>= 16; r += 16; }
470 if (x >> 8) { x >>= 8; r += 8; } 535 if (x >> 8) { x >>= 8; r += 8; }
471 if (x >> 4) { x >>= 4; r += 4; } 536 if (x >> 4) { x >>= 4; r += 4; }
472 if (x >> 2) { x >>= 2; r += 2; } 537 if (x >> 2) { x >>= 2; r += 2; }
473 if (x >> 1) { r += 1; } 538 if (x >> 1) { r += 1; }
474 539
475 return r; 540 return r;
541#endif
476 } 542 }
477 543
478 ecb_function_ ecb_const int ecb_ld64 (uint64_t x); 544 ecb_function_ ecb_const int ecb_ld64 (uint64_t x);
479 ecb_function_ ecb_const int ecb_ld64 (uint64_t x) 545 ecb_function_ ecb_const int ecb_ld64 (uint64_t x)
480 { 546 {
547#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
548 unsigned long r;
549 _BitScanReverse64 (&r, x);
550 return (int)r;
551#else
481 int r = 0; 552 int r = 0;
482 553
483 if (x >> 32) { x >>= 32; r += 32; } 554 if (x >> 32) { x >>= 32; r += 32; }
484 555
485 return r + ecb_ld32 (x); 556 return r + ecb_ld32 (x);
557#endif
486 } 558 }
487#endif 559#endif
488 560
489ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x); 561ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x);
490ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); } 562ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); }
… …
537ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count); 609ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count);
538ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count); 610ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count);
539ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count); 611ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count);
540ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count); 612ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count);
541 613
542ecb_inline ecb_const uint8_t ecb_rotl8 (uint8_t x, unsigned int count) { return (x >> ( 8 - count)) | (x << count); } 614ecb_inline ecb_const uint8_t ecb_rotl8 (uint8_t x, unsigned int count) { return (x >> (-count & 7)) | (x << (count & 7)); }
543ecb_inline ecb_const uint8_t ecb_rotr8 (uint8_t x, unsigned int count) { return (x << ( 8 - count)) | (x >> count); } 615ecb_inline ecb_const uint8_t ecb_rotr8 (uint8_t x, unsigned int count) { return (x << (-count & 7)) | (x >> (count & 7)); }
544ecb_inline ecb_const uint16_t ecb_rotl16 (uint16_t x, unsigned int count) { return (x >> (16 - count)) | (x << count); } 616ecb_inline ecb_const uint16_t ecb_rotl16 (uint16_t x, unsigned int count) { return (x >> (-count & 15)) | (x << (count & 15)); }
545ecb_inline ecb_const uint16_t ecb_rotr16 (uint16_t x, unsigned int count) { return (x << (16 - count)) | (x >> count); } 617ecb_inline ecb_const uint16_t ecb_rotr16 (uint16_t x, unsigned int count) { return (x << (-count & 15)) | (x >> (count & 15)); }
546ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); } 618ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (-count & 31)) | (x << (count & 31)); }
547ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); } 619ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (-count & 31)) | (x >> (count & 31)); }
548ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); } 620ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (-count & 63)) | (x << (count & 63)); }
549ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); } 621ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (-count & 63)) | (x >> (count & 63)); }
622
623#if ECB_CPP
624
625inline uint8_t ecb_ctz (uint8_t v) { return ecb_ctz32 (v); }
626inline uint16_t ecb_ctz (uint16_t v) { return ecb_ctz32 (v); }
627inline uint32_t ecb_ctz (uint32_t v) { return ecb_ctz32 (v); }
628inline uint64_t ecb_ctz (uint64_t v) { return ecb_ctz64 (v); }
629
630inline bool ecb_is_pot (uint8_t v) { return ecb_is_pot32 (v); }
631inline bool ecb_is_pot (uint16_t v) { return ecb_is_pot32 (v); }
632inline bool ecb_is_pot (uint32_t v) { return ecb_is_pot32 (v); }
633inline bool ecb_is_pot (uint64_t v) { return ecb_is_pot64 (v); }
634
635inline int ecb_ld (uint8_t v) { return ecb_ld32 (v); }
636inline int ecb_ld (uint16_t v) { return ecb_ld32 (v); }
637inline int ecb_ld (uint32_t v) { return ecb_ld32 (v); }
638inline int ecb_ld (uint64_t v) { return ecb_ld64 (v); }
639
640inline int ecb_popcount (uint8_t v) { return ecb_popcount32 (v); }
641inline int ecb_popcount (uint16_t v) { return ecb_popcount32 (v); }
642inline int ecb_popcount (uint32_t v) { return ecb_popcount32 (v); }
643inline int ecb_popcount (uint64_t v) { return ecb_popcount64 (v); }
644
645inline uint8_t ecb_bitrev (uint8_t v) { return ecb_bitrev8 (v); }
646inline uint16_t ecb_bitrev (uint16_t v) { return ecb_bitrev16 (v); }
647inline uint32_t ecb_bitrev (uint32_t v) { return ecb_bitrev32 (v); }
648
649inline uint8_t ecb_rotl (uint8_t v, unsigned int count) { return ecb_rotl8 (v, count); }
650inline uint16_t ecb_rotl (uint16_t v, unsigned int count) { return ecb_rotl16 (v, count); }
651inline uint32_t ecb_rotl (uint32_t v, unsigned int count) { return ecb_rotl32 (v, count); }
652inline uint64_t ecb_rotl (uint64_t v, unsigned int count) { return ecb_rotl64 (v, count); }
653
654inline uint8_t ecb_rotr (uint8_t v, unsigned int count) { return ecb_rotr8 (v, count); }
655inline uint16_t ecb_rotr (uint16_t v, unsigned int count) { return ecb_rotr16 (v, count); }
656inline uint32_t ecb_rotr (uint32_t v, unsigned int count) { return ecb_rotr32 (v, count); }
657inline uint64_t ecb_rotr (uint64_t v, unsigned int count) { return ecb_rotr64 (v, count); }
658
659#endif
550 660
551#if ECB_GCC_VERSION(4,3) || (ECB_CLANG_BUILTIN(__builtin_bswap32) && ECB_CLANG_BUILTIN(__builtin_bswap64)) 661#if ECB_GCC_VERSION(4,3) || (ECB_CLANG_BUILTIN(__builtin_bswap32) && ECB_CLANG_BUILTIN(__builtin_bswap64))
662 #if ECB_GCC_VERSION(4,8) || ECB_CLANG_BUILTIN(__builtin_bswap16)
663 #define ecb_bswap16(x) __builtin_bswap16 (x)
664 #else
552 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16) 665 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16)
666 #endif
553 #define ecb_bswap32(x) __builtin_bswap32 (x) 667 #define ecb_bswap32(x) __builtin_bswap32 (x)
554 #define ecb_bswap64(x) __builtin_bswap64 (x) 668 #define ecb_bswap64(x) __builtin_bswap64 (x)
669#elif _MSC_VER
670 #include <stdlib.h>
671 #define ecb_bswap16(x) ((uint16_t)_byteswap_ushort ((uint16_t)(x)))
672 #define ecb_bswap32(x) ((uint32_t)_byteswap_ulong ((uint32_t)(x)))
673 #define ecb_bswap64(x) ((uint64_t)_byteswap_uint64 ((uint64_t)(x)))
555#else 674#else
556 ecb_function_ ecb_const uint16_t ecb_bswap16 (uint16_t x); 675 ecb_function_ ecb_const uint16_t ecb_bswap16 (uint16_t x);
557 ecb_function_ ecb_const uint16_t 676 ecb_function_ ecb_const uint16_t
558 ecb_bswap16 (uint16_t x) 677 ecb_bswap16 (uint16_t x)
559 { 678 {
… …
584#endif 703#endif
585 704
586/* try to tell the compiler that some condition is definitely true */ 705/* try to tell the compiler that some condition is definitely true */
587#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0 706#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0
588 707
589ecb_inline ecb_const unsigned char ecb_byteorder_helper (void); 708ecb_inline ecb_const uint32_t ecb_byteorder_helper (void);
590ecb_inline ecb_const unsigned char 709ecb_inline ecb_const uint32_t
591ecb_byteorder_helper (void) 710ecb_byteorder_helper (void)
592{ 711{
593 /* the union code still generates code under pressure in gcc, */ 712 /* the union code still generates code under pressure in gcc, */
594 /* but less than using pointers, and always seems to */ 713 /* but less than using pointers, and always seems to */
595 /* successfully return a constant. */ 714 /* successfully return a constant. */
596 /* the reason why we have this horrible preprocessor mess */ 715 /* the reason why we have this horrible preprocessor mess */
597 /* is to avoid it in all cases, at least on common architectures */ 716 /* is to avoid it in all cases, at least on common architectures */
598 /* or when using a recent enough gcc version (>= 4.6) */ 717 /* or when using a recent enough gcc version (>= 4.6) */
599#if ((__i386 || __i386__) && !__VOS__) || _M_IX86 || ECB_GCC_AMD64 || ECB_MSVC_AMD64
600 return 0x44;
601#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ 718#if (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) \
719 || ((__i386 || __i386__ || _M_IX86 || ECB_GCC_AMD64 || ECB_MSVC_AMD64) && !__VOS__)
720 #define ECB_LITTLE_ENDIAN 1
602 return 0x44; 721 return 0x44332211;
603#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ 722#elif (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) \
723 || ((__AARCH64EB__ || __MIPSEB__ || __ARMEB__) && !__VOS__)
724 #define ECB_BIG_ENDIAN 1
604 return 0x11; 725 return 0x11223344;
605#else 726#else
606 union 727 union
607 { 728 {
729 uint8_t c[4];
608 uint32_t i; 730 uint32_t u;
609 uint8_t c;
610 } u = { 0x11223344 }; 731 } u = { 0x11, 0x22, 0x33, 0x44 };
611 return u.c; 732 return u.u;
612#endif 733#endif
613} 734}
614 735
615ecb_inline ecb_const ecb_bool ecb_big_endian (void); 736ecb_inline ecb_const ecb_bool ecb_big_endian (void);
616ecb_inline ecb_const ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; } 737ecb_inline ecb_const ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11223344; }
617ecb_inline ecb_const ecb_bool ecb_little_endian (void); 738ecb_inline ecb_const ecb_bool ecb_little_endian (void);
618ecb_inline ecb_const ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; } 739ecb_inline ecb_const ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44332211; }
740
741/*****************************************************************************/
742/* unaligned load/store */
743
744ecb_inline uint_fast16_t ecb_be_u16_to_host (uint_fast16_t v) { return ecb_little_endian () ? ecb_bswap16 (v) : v; }
745ecb_inline uint_fast32_t ecb_be_u32_to_host (uint_fast32_t v) { return ecb_little_endian () ? ecb_bswap32 (v) : v; }
746ecb_inline uint_fast64_t ecb_be_u64_to_host (uint_fast64_t v) { return ecb_little_endian () ? ecb_bswap64 (v) : v; }
747
748ecb_inline uint_fast16_t ecb_le_u16_to_host (uint_fast16_t v) { return ecb_big_endian () ? ecb_bswap16 (v) : v; }
749ecb_inline uint_fast32_t ecb_le_u32_to_host (uint_fast32_t v) { return ecb_big_endian () ? ecb_bswap32 (v) : v; }
750ecb_inline uint_fast64_t ecb_le_u64_to_host (uint_fast64_t v) { return ecb_big_endian () ? ecb_bswap64 (v) : v; }
751
752ecb_inline uint_fast16_t ecb_peek_u16_u (const void *ptr) { uint16_t v; memcpy (&v, ptr, sizeof (v)); return v; }
753ecb_inline uint_fast32_t ecb_peek_u32_u (const void *ptr) { uint32_t v; memcpy (&v, ptr, sizeof (v)); return v; }
754ecb_inline uint_fast64_t ecb_peek_u64_u (const void *ptr) { uint64_t v; memcpy (&v, ptr, sizeof (v)); return v; }
755
756ecb_inline uint_fast16_t ecb_peek_be_u16_u (const void *ptr) { return ecb_be_u16_to_host (ecb_peek_u16_u (ptr)); }
757ecb_inline uint_fast32_t ecb_peek_be_u32_u (const void *ptr) { return ecb_be_u32_to_host (ecb_peek_u32_u (ptr)); }
758ecb_inline uint_fast64_t ecb_peek_be_u64_u (const void *ptr) { return ecb_be_u64_to_host (ecb_peek_u64_u (ptr)); }
759
760ecb_inline uint_fast16_t ecb_peek_le_u16_u (const void *ptr) { return ecb_le_u16_to_host (ecb_peek_u16_u (ptr)); }
761ecb_inline uint_fast32_t ecb_peek_le_u32_u (const void *ptr) { return ecb_le_u32_to_host (ecb_peek_u32_u (ptr)); }
762ecb_inline uint_fast64_t ecb_peek_le_u64_u (const void *ptr) { return ecb_le_u64_to_host (ecb_peek_u64_u (ptr)); }
763
764ecb_inline uint_fast16_t ecb_host_to_be_u16 (uint_fast16_t v) { return ecb_little_endian () ? ecb_bswap16 (v) : v; }
765ecb_inline uint_fast32_t ecb_host_to_be_u32 (uint_fast32_t v) { return ecb_little_endian () ? ecb_bswap32 (v) : v; }
766ecb_inline uint_fast64_t ecb_host_to_be_u64 (uint_fast64_t v) { return ecb_little_endian () ? ecb_bswap64 (v) : v; }
767
768ecb_inline uint_fast16_t ecb_host_to_le_u16 (uint_fast16_t v) { return ecb_big_endian () ? ecb_bswap16 (v) : v; }
769ecb_inline uint_fast32_t ecb_host_to_le_u32 (uint_fast32_t v) { return ecb_big_endian () ? ecb_bswap32 (v) : v; }
770ecb_inline uint_fast64_t ecb_host_to_le_u64 (uint_fast64_t v) { return ecb_big_endian () ? ecb_bswap64 (v) : v; }
771
772ecb_inline void ecb_poke_u16_u (void *ptr, uint16_t v) { memcpy (ptr, &v, sizeof (v)); }
773ecb_inline void ecb_poke_u32_u (void *ptr, uint32_t v) { memcpy (ptr, &v, sizeof (v)); }
774ecb_inline void ecb_poke_u64_u (void *ptr, uint64_t v) { memcpy (ptr, &v, sizeof (v)); }
775
776ecb_inline void ecb_poke_be_u16_u (void *ptr, uint_fast16_t v) { ecb_poke_u16_u (ptr, ecb_host_to_be_u16 (v)); }
777ecb_inline void ecb_poke_be_u32_u (void *ptr, uint_fast32_t v) { ecb_poke_u32_u (ptr, ecb_host_to_be_u32 (v)); }
778ecb_inline void ecb_poke_be_u64_u (void *ptr, uint_fast64_t v) { ecb_poke_u64_u (ptr, ecb_host_to_be_u64 (v)); }
779
780ecb_inline void ecb_poke_le_u16_u (void *ptr, uint_fast16_t v) { ecb_poke_u16_u (ptr, ecb_host_to_le_u16 (v)); }
781ecb_inline void ecb_poke_le_u32_u (void *ptr, uint_fast32_t v) { ecb_poke_u32_u (ptr, ecb_host_to_le_u32 (v)); }
782ecb_inline void ecb_poke_le_u64_u (void *ptr, uint_fast64_t v) { ecb_poke_u64_u (ptr, ecb_host_to_le_u64 (v)); }
783
784#if ECB_CPP
785
786inline uint8_t ecb_bswap (uint8_t v) { return v; }
787inline uint16_t ecb_bswap (uint16_t v) { return ecb_bswap16 (v); }
788inline uint32_t ecb_bswap (uint32_t v) { return ecb_bswap32 (v); }
789inline uint64_t ecb_bswap (uint64_t v) { return ecb_bswap64 (v); }
790
791template<typename T> inline T ecb_be_to_host (T v) { return ecb_little_endian () ? ecb_bswap (v) : v; }
792template<typename T> inline T ecb_le_to_host (T v) { return ecb_big_endian () ? ecb_bswap (v) : v; }
793template<typename T> inline T ecb_peek (const void *ptr) { return *(const T *)ptr; }
794template<typename T> inline T ecb_peek_be (const void *ptr) { return ecb_be_to_host (ecb_peek <T> (ptr)); }
795template<typename T> inline T ecb_peek_le (const void *ptr) { return ecb_le_to_host (ecb_peek <T> (ptr)); }
796template<typename T> inline T ecb_peek_u (const void *ptr) { T v; memcpy (&v, ptr, sizeof (v)); return v; }
797template<typename T> inline T ecb_peek_be_u (const void *ptr) { return ecb_be_to_host (ecb_peek_u<T> (ptr)); }
798template<typename T> inline T ecb_peek_le_u (const void *ptr) { return ecb_le_to_host (ecb_peek_u<T> (ptr)); }
799
800template<typename T> inline T ecb_host_to_be (T v) { return ecb_little_endian () ? ecb_bswap (v) : v; }
801template<typename T> inline T ecb_host_to_le (T v) { return ecb_big_endian () ? ecb_bswap (v) : v; }
802template<typename T> inline void ecb_poke (void *ptr, T v) { *(T *)ptr = v; }
803template<typename T> inline void ecb_poke_be (void *ptr, T v) { return ecb_poke <T> (ptr, ecb_host_to_be (v)); }
804template<typename T> inline void ecb_poke_le (void *ptr, T v) { return ecb_poke <T> (ptr, ecb_host_to_le (v)); }
805template<typename T> inline void ecb_poke_u (void *ptr, T v) { memcpy (ptr, &v, sizeof (v)); }
806template<typename T> inline void ecb_poke_be_u (void *ptr, T v) { return ecb_poke_u<T> (ptr, ecb_host_to_be (v)); }
807template<typename T> inline void ecb_poke_le_u (void *ptr, T v) { return ecb_poke_u<T> (ptr, ecb_host_to_le (v)); }
808
809#endif
810
811/*****************************************************************************/
812/* division */
619 813
620#if ECB_GCC_VERSION(3,0) || ECB_C99 814#if ECB_GCC_VERSION(3,0) || ECB_C99
815 /* C99 tightened the definition of %, so we can use a more efficient version */
621 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) 816 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0))
622#else 817#else
623 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) 818 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n)))
624#endif 819#endif
625 820
… …
636 } 831 }
637#else 832#else
638 #define ecb_div_rd(val,div) ((val) < 0 ? - ((-(val) + (div) - 1) / (div)) : ((val) ) / (div)) 833 #define ecb_div_rd(val,div) ((val) < 0 ? - ((-(val) + (div) - 1) / (div)) : ((val) ) / (div))
639 #define ecb_div_ru(val,div) ((val) < 0 ? - ((-(val) ) / (div)) : ((val) + (div) - 1) / (div)) 834 #define ecb_div_ru(val,div) ((val) < 0 ? - ((-(val) ) / (div)) : ((val) + (div) - 1) / (div))
640#endif 835#endif
836
837/*****************************************************************************/
838/* array length */
641 839
642#if ecb_cplusplus_does_not_suck 840#if ecb_cplusplus_does_not_suck
643 /* does not work for local types (http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2008/n2657.htm) */ 841 /* does not work for local types (http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2008/n2657.htm) */
644 template<typename T, int N> 842 template<typename T, int N>
645 static inline int ecb_array_length (const T (&arr)[N]) 843 static inline int ecb_array_length (const T (&arr)[N])
… …
647 return N; 845 return N;
648 } 846 }
649#else 847#else
650 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) 848 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0]))
651#endif 849#endif
850
851/*****************************************************************************/
852/* IEEE 754-2008 half float conversions */
853
854ecb_function_ ecb_const uint32_t ecb_binary16_to_binary32 (uint32_t x);
855ecb_function_ ecb_const uint32_t
856ecb_binary16_to_binary32 (uint32_t x)
857{
858 unsigned int s = (x & 0x8000) << (31 - 15);
859 int e = (x >> 10) & 0x001f;
860 unsigned int m = x & 0x03ff;
861
862 if (ecb_expect_false (e == 31))
863 /* infinity or NaN */
864 e = 255 - (127 - 15);
865 else if (ecb_expect_false (!e))
866 {
867 if (ecb_expect_true (!m))
868 /* zero, handled by code below by forcing e to 0 */
869 e = 0 - (127 - 15);
870 else
871 {
872 /* subnormal, renormalise */
873 unsigned int s = 10 - ecb_ld32 (m);
874
875 m = (m << s) & 0x3ff; /* mask implicit bit */
876 e -= s - 1;
877 }
878 }
879
880 /* e and m now are normalised, or zero, (or inf or nan) */
881 e += 127 - 15;
882
883 return s | (e << 23) | (m << (23 - 10));
884}
885
886ecb_function_ ecb_const uint16_t ecb_binary32_to_binary16 (uint32_t x);
887ecb_function_ ecb_const uint16_t
888ecb_binary32_to_binary16 (uint32_t x)
889{
890 unsigned int s = (x >> 16) & 0x00008000; /* sign bit, the easy part */
891 int e = ((x >> 23) & 0x000000ff) - (127 - 15); /* the desired exponent */
892 unsigned int m = x & 0x007fffff;
893
894 x &= 0x7fffffff;
895
896 /* if it's within range of binary16 normals, use fast path */
897 if (ecb_expect_true (0x38800000 <= x && x <= 0x477fefff))
898 {
899 /* mantissa round-to-even */
900 m += 0x00000fff + ((m >> (23 - 10)) & 1);
901
902 /* handle overflow */
903 if (ecb_expect_false (m >= 0x00800000))
904 {
905 m >>= 1;
906 e += 1;
907 }
908
909 return s | (e << 10) | (m >> (23 - 10));
910 }
911
912 /* handle large numbers and infinity */
913 if (ecb_expect_true (0x477fefff < x && x <= 0x7f800000))
914 return s | 0x7c00;
915
916 /* handle zero, subnormals and small numbers */
917 if (ecb_expect_true (x < 0x38800000))
918 {
919 /* zero */
920 if (ecb_expect_true (!x))
921 return s;
922
923 /* handle subnormals */
924
925 /* too small, will be zero */
926 if (e < (14 - 24)) /* might not be sharp, but is good enough */
927 return s;
928
929 m |= 0x00800000; /* make implicit bit explicit */
930
931 /* very tricky - we need to round to the nearest e (+10) bit value */
932 {
933 unsigned int bits = 14 - e;
934 unsigned int half = (1 << (bits - 1)) - 1;
935 unsigned int even = (m >> bits) & 1;
936
937 /* if this overflows, we will end up with a normalised number */
938 m = (m + half + even) >> bits;
939 }
940
941 return s | m;
942 }
943
944 /* handle NaNs, preserve leftmost nan bits, but make sure we don't turn them into infinities */
945 m >>= 13;
946
947 return s | 0x7c00 | m | !m;
948}
949
950/*******************************************************************************/
951/* fast integer to ascii */
952
953/*
954 * This code is pretty complicated because it is general. The idea behind it,
955 * however, is pretty simple: first, the number is multiplied with a scaling
956 * factor (2**bits / 10**(digits-1)) to convert the integer into a fixed-point
957 * number with the first digit in the upper bits.
958 * Then this digit is converted to text and masked out. The resulting number
959 * is then multiplied by 10, by multiplying the fixed point representation
960 * by 5 and shifting the (binary) decimal point one to the right, so a 4.28
961 * format becomes 5.27, 6.26 and so on.
962 * The rest involves only advancing the pointer if we already generated a
963 * non-zero digit, so leading zeroes are overwritten.
964 */
965
966// simply return a mask with "bits" bits set
967#define ecb_i2a_mask(type,bits) ((((type)1) << (bits)) - 1)
968
969// oputput a single digit. maskvalue is 10**digitidx
970#define ecb_i2a_digit(type,bits,digitmask,maskvalue,digitidx) \
971 if (digitmask >= maskvalue) /* constant, used to decide how many digits to generate */ \
972 { \
973 char digit = x >> (bits - digitidx); /* calculate the topmost digit */ \
974 *ptr = digit + '0'; /* output it */ \
975 nz = (digitmask == maskvalue) || nz || digit; /* first term == always output last digit */ \
976 ptr += nz; /* output digit only if non-zero digit seen */ \
977 x = (x & ecb_i2a_mask (type, bits - digitidx)) * 5; /* *10, but shift decimal point right */ \
978 }
979
980// convert integer to fixed point format and multiply out digits, highest first
981// requires magic constants: max. digits and number of bits after the decimal point
982#define ecb_i2a_def(suffix,ptr,v,type,bits,digitmask,lz) \
983ecb_inline char *ecb_i2a_ ## suffix (char *ptr, uint32_t u) \
984{ \
985 char nz = lz; /* non-zero digit seen? */ \
986 /* convert to x.bits fixed-point */ \
987 type x = u * ((ecb_i2a_mask (type, bits) + digitmask) / digitmask); \
988 /* output up to 10 digits */ \
989 ecb_i2a_digit (type,bits,digitmask, 1, 0); \
990 ecb_i2a_digit (type,bits,digitmask, 10, 1); \
991 ecb_i2a_digit (type,bits,digitmask, 100, 2); \
992 ecb_i2a_digit (type,bits,digitmask, 1000, 3); \
993 ecb_i2a_digit (type,bits,digitmask, 10000, 4); \
994 ecb_i2a_digit (type,bits,digitmask, 100000, 5); \
995 ecb_i2a_digit (type,bits,digitmask, 1000000, 6); \
996 ecb_i2a_digit (type,bits,digitmask, 10000000, 7); \
997 ecb_i2a_digit (type,bits,digitmask, 100000000, 8); \
998 ecb_i2a_digit (type,bits,digitmask, 1000000000, 9); \
999 return ptr; \
1000}
1001
1002// predefined versions of the above, for various digits
1003// ecb_i2a_xN = almost N digits, limit defined by macro
1004// ecb_i2a_N = up to N digits, leading zeroes suppressed
1005// ecb_i2a_0N = exactly N digits, including leading zeroes
1006
1007// non-leading-zero versions, limited range
1008#define ECB_I2A_MAX_X5 59074 // limit for ecb_i2a_x5
1009#define ECB_I2A_MAX_X10 2932500665 // limit for ecb_i2a_x10
1010ecb_i2a_def ( x5, ptr, v, uint32_t, 26, 10000, 0)
1011ecb_i2a_def (x10, ptr, v, uint64_t, 60, 1000000000, 0)
1012
1013// non-leading zero versions, all digits, 4 and 9 are optimal for 32/64 bit
1014ecb_i2a_def ( 2, ptr, v, uint32_t, 10, 10, 0)
1015ecb_i2a_def ( 3, ptr, v, uint32_t, 12, 100, 0)
1016ecb_i2a_def ( 4, ptr, v, uint32_t, 26, 1000, 0)
1017ecb_i2a_def ( 5, ptr, v, uint64_t, 30, 10000, 0)
1018ecb_i2a_def ( 6, ptr, v, uint64_t, 36, 100000, 0)
1019ecb_i2a_def ( 7, ptr, v, uint64_t, 44, 1000000, 0)
1020ecb_i2a_def ( 8, ptr, v, uint64_t, 50, 10000000, 0)
1021ecb_i2a_def ( 9, ptr, v, uint64_t, 56, 100000000, 0)
1022
1023// leading-zero versions, all digits, 04 and 09 are optimal for 32/64 bit
1024ecb_i2a_def (02, ptr, v, uint32_t, 10, 10, 1)
1025ecb_i2a_def (03, ptr, v, uint32_t, 12, 100, 1)
1026ecb_i2a_def (04, ptr, v, uint32_t, 26, 1000, 1)
1027ecb_i2a_def (05, ptr, v, uint64_t, 30, 10000, 1)
1028ecb_i2a_def (06, ptr, v, uint64_t, 36, 100000, 1)
1029ecb_i2a_def (07, ptr, v, uint64_t, 44, 1000000, 1)
1030ecb_i2a_def (08, ptr, v, uint64_t, 50, 10000000, 1)
1031ecb_i2a_def (09, ptr, v, uint64_t, 56, 100000000, 1)
1032
1033#define ECB_I2A_I32_DIGITS 11
1034#define ECB_I2A_U32_DIGITS 10
1035#define ECB_I2A_I64_DIGITS 20
1036#define ECB_I2A_U64_DIGITS 21
1037#define ECB_I2A_MAX_DIGITS 21
1038
1039ecb_inline char *
1040ecb_i2a_u32 (char *ptr, uint32_t u)
1041{
1042 #if ECB_64BIT_NATIVE
1043 if (ecb_expect_true (u <= ECB_I2A_MAX_X10))
1044 ptr = ecb_i2a_x10 (ptr, u);
1045 else // x10 almost, but not fully, covers 32 bit
1046 {
1047 uint32_t u1 = u % 1000000000;
1048 uint32_t u2 = u / 1000000000;
1049
1050 *ptr++ = u2 + '0';
1051 ptr = ecb_i2a_09 (ptr, u1);
1052 }
1053 #else
1054 if (ecb_expect_true (u <= ECB_I2A_MAX_X5))
1055 ecb_i2a_x5 (ptr, u);
1056 else if (ecb_expect_true (u <= ECB_I2A_MAX_X5 * 10000))
1057 {
1058 uint32_t u1 = u % 10000;
1059 uint32_t u2 = u / 10000;
1060
1061 ptr = ecb_i2a_x5 (ptr, u2);
1062 ptr = ecb_i2a_04 (ptr, u1);
1063 }
1064 else
1065 {
1066 uint32_t u1 = u % 10000;
1067 uint32_t ua = u / 10000;
1068 uint32_t u2 = ua % 10000;
1069 uint32_t u3 = ua / 10000;
1070
1071 ptr = ecb_i2a_2 (ptr, u3);
1072 ptr = ecb_i2a_04 (ptr, u2);
1073 ptr = ecb_i2a_04 (ptr, u1);
1074 }
1075 #endif
1076
1077 return ptr;
1078}
1079
1080ecb_inline char *
1081ecb_i2a_i32 (char *ptr, int32_t v)
1082{
1083 *ptr = '-'; ptr += v < 0;
1084 uint32_t u = v < 0 ? -(uint32_t)v : v;
1085
1086 #if ECB_64BIT_NATIVE
1087 ptr = ecb_i2a_x10 (ptr, u); // x10 fully covers 31 bit
1088 #else
1089 ptr = ecb_i2a_u32 (ptr, u);
1090 #endif
1091
1092 return ptr;
1093}
1094
1095ecb_inline char *
1096ecb_i2a_u64 (char *ptr, uint64_t u)
1097{
1098 #if ECB_64BIT_NATIVE
1099 if (ecb_expect_true (u <= ECB_I2A_MAX_X10))
1100 ptr = ecb_i2a_x10 (ptr, u);
1101 else if (ecb_expect_false (u <= ECB_I2A_MAX_X10 * 1000000000))
1102 {
1103 uint64_t u1 = u % 1000000000;
1104 uint64_t u2 = u / 1000000000;
1105
1106 ptr = ecb_i2a_x10 (ptr, u2);
1107 ptr = ecb_i2a_09 (ptr, u1);
1108 }
1109 else
1110 {
1111 uint64_t u1 = u % 1000000000;
1112 uint64_t ua = u / 1000000000;
1113 uint64_t u2 = ua % 1000000000;
1114 uint64_t u3 = ua / 1000000000;
1115
1116 ptr = ecb_i2a_2 (ptr, u3);
1117 ptr = ecb_i2a_09 (ptr, u2);
1118 ptr = ecb_i2a_09 (ptr, u1);
1119 }
1120 #else
1121 if (ecb_expect_true (u <= ECB_I2A_MAX_X5))
1122 ptr = ecb_i2a_x5 (ptr, u);
1123 else
1124 {
1125 uint64_t u1 = u % 10000;
1126 uint64_t u2 = u / 10000;
1127
1128 ptr = ecb_i2a_u64 (ptr, u2);
1129 ptr = ecb_i2a_04 (ptr, u1);
1130 }
1131 #endif
1132
1133 return ptr;
1134}
1135
1136ecb_inline char *
1137ecb_i2a_i64 (char *ptr, int64_t v)
1138{
1139 *ptr = '-'; ptr += v < 0;
1140 uint64_t u = v < 0 ? -(uint64_t)v : v;
1141
1142 #if ECB_64BIT_NATIVE
1143 if (ecb_expect_true (u <= ECB_I2A_MAX_X10))
1144 ptr = ecb_i2a_x10 (ptr, u);
1145 else if (ecb_expect_false (u <= ECB_I2A_MAX_X10 * 1000000000))
1146 {
1147 uint64_t u1 = u % 1000000000;
1148 uint64_t u2 = u / 1000000000;
1149
1150 ptr = ecb_i2a_x10 (ptr, u2);
1151 ptr = ecb_i2a_09 (ptr, u1);
1152 }
1153 else
1154 {
1155 uint64_t u1 = u % 1000000000;
1156 uint64_t ua = u / 1000000000;
1157 uint64_t u2 = ua % 1000000000;
1158 uint64_t u3 = ua / 1000000000;
1159
1160 // 2**31 is 19 digits, so the top is exactly one digit
1161 *ptr++ = u3 + '0';
1162 ptr = ecb_i2a_09 (ptr, u2);
1163 ptr = ecb_i2a_09 (ptr, u1);
1164 }
1165 #else
1166 ptr = ecb_i2a_u64 (ptr, u);
1167 #endif
1168
1169 return ptr;
1170}
652 1171
653/*******************************************************************************/ 1172/*******************************************************************************/
654/* floating point stuff, can be disabled by defining ECB_NO_LIBM */ 1173/* floating point stuff, can be disabled by defining ECB_NO_LIBM */
655 1174
656/* basically, everything uses "ieee pure-endian" floating point numbers */ 1175/* basically, everything uses "ieee pure-endian" floating point numbers */
… …
669 || defined __sh__ \ 1188 || defined __sh__ \
670 || defined _M_IX86 || defined ECB_MSVC_AMD64 || defined _M_IA64 \ 1189 || defined _M_IX86 || defined ECB_MSVC_AMD64 || defined _M_IA64 \
671 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \ 1190 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \
672 || defined __aarch64__ 1191 || defined __aarch64__
673 #define ECB_STDFP 1 1192 #define ECB_STDFP 1
674 #include <string.h> /* for memcpy */
675#else 1193#else
676 #define ECB_STDFP 0 1194 #define ECB_STDFP 0
677#endif 1195#endif
678 1196
679#ifndef ECB_NO_LIBM 1197#ifndef ECB_NO_LIBM
… …
693 #define ECB_NAN ECB_INFINITY 1211 #define ECB_NAN ECB_INFINITY
694 #endif 1212 #endif
695 1213
696 #if ECB_C99 || _XOPEN_VERSION >= 600 || _POSIX_VERSION >= 200112L 1214 #if ECB_C99 || _XOPEN_VERSION >= 600 || _POSIX_VERSION >= 200112L
697 #define ecb_ldexpf(x,e) ldexpf ((x), (e)) 1215 #define ecb_ldexpf(x,e) ldexpf ((x), (e))
1216 #define ecb_frexpf(x,e) frexpf ((x), (e))
698 #else 1217 #else
699 #define ecb_ldexpf(x,e) (float) ldexp ((x), (e)) 1218 #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e))
1219 #define ecb_frexpf(x,e) (float) frexp ((double) (x), (e))
700 #endif 1220 #endif
701
702 /* converts an ieee half/binary16 to a float */
703 ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
704 ecb_function_ ecb_const float
705 ecb_binary16_to_float (uint16_t x)
706 {
707 int e = (x >> 10) & 0x1f;
708 int m = x & 0x3ff;
709 float r;
710
711 if (!e ) r = ecb_ldexpf (m , -24);
712 else if (e != 31) r = ecb_ldexpf (m + 0x400, e - 25);
713 else if (m ) r = ECB_NAN;
714 else r = ECB_INFINITY;
715
716 return x & 0x8000 ? -r : r;
717 }
718 1221
719 /* convert a float to ieee single/binary32 */ 1222 /* convert a float to ieee single/binary32 */
720 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x); 1223 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x);
721 ecb_function_ ecb_const uint32_t 1224 ecb_function_ ecb_const uint32_t
722 ecb_float_to_binary32 (float x) 1225 ecb_float_to_binary32 (float x)
… …
733 if (x == 0e0f ) return 0x00000000U; 1236 if (x == 0e0f ) return 0x00000000U;
734 if (x > +3.40282346638528860e+38f) return 0x7f800000U; 1237 if (x > +3.40282346638528860e+38f) return 0x7f800000U;
735 if (x < -3.40282346638528860e+38f) return 0xff800000U; 1238 if (x < -3.40282346638528860e+38f) return 0xff800000U;
736 if (x != x ) return 0x7fbfffffU; 1239 if (x != x ) return 0x7fbfffffU;
737 1240
738 m = frexpf (x, &e) * 0x1000000U; 1241 m = ecb_frexpf (x, &e) * 0x1000000U;
739 1242
740 r = m & 0x80000000U; 1243 r = m & 0x80000000U;
741 1244
742 if (r) 1245 if (r)
743 m = -m; 1246 m = -m;
… …
854 #endif 1357 #endif
855 1358
856 return r; 1359 return r;
857 } 1360 }
858 1361
859#endif 1362 /* convert a float to ieee half/binary16 */
1363 ecb_function_ ecb_const uint16_t ecb_float_to_binary16 (float x);
1364 ecb_function_ ecb_const uint16_t
1365 ecb_float_to_binary16 (float x)
1366 {
1367 return ecb_binary32_to_binary16 (ecb_float_to_binary32 (x));
1368 }
860 1369
861#endif 1370 /* convert an ieee half/binary16 to float */
1371 ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
1372 ecb_function_ ecb_const float
1373 ecb_binary16_to_float (uint16_t x)
1374 {
1375 return ecb_binary32_to_float (ecb_binary16_to_binary32 (x));
1376 }
862 1377
1378#endif
1379
1380#endif
1381

Diff Legend

– Removed lines
+ Added lines
< Changed lines
> Changed lines